Novel segmentation algorithm for high-throughput analysis of spectral domain-optical coherence tomography imaging of teleost retina

Aim Spectral Domain-Optical Coherence Tomography (SD-OCT) has become an essential tool to assess the health of ocular tissues in live subjects. The processing of SD-OCT images, in particular from non-mammalian species, is a labour-intensive manual process due to the lack of access to analytical programs. The work presented herein describes the development and implementation of a novel computer algorithm for quantitative analysis of SD-OCT images of live teleost eyes. We hypothesized that this algorithm, in comparison to manual segmentation of SD-OCT images, will allow more precise measurement, with significantly higher throughput capacity, of retinal architecture in live teleost ocular tissue. Methods Automated segmentation processing of SD-OCT images of the retinal layers was developed using a novel algorithm based on thresholding, which operates on the pixel values contained in an image. The algorithm measured the thickness characteristic of the retina present in the input dataset to provide increased accuracy and repeatability of SD-OCT analysis over manual measurements. The program was also designed to allow adjustments of the thresholding variables to suit any specific image set. A heat map software was created alongside the algorithm to plot the SD-OCT image measurements as a colour gradient. Results Automated segmentation analysis of the retinal layers from SD-OCT images enabled analysis of a large volume of imaging data of teleost ocular structures in a short time. The algorithm was just as accurate when compared to manual measurements and provided repeatability as measurements could be quickly reassessed to confirm previously determined results. This is the case as the algorithm can generate hundreds of retinal thickness measurements per image for a large number of images for a given dataset. This algorithm can be deemed as repeatable as each input will always produce the same output due to the thresholding methods used. This is the case as thresholding is a finite mathematical process to determine a range of values. The measurements produced from this assessment were represented by a heat map software that directly converted the measurements taken from each processed image to represent the changes in thickness across the whole retinal scan. Conclusions Our work addresses the need for accurate and high-throughput SDOCT data analysis for the retinal tissues of teleosts where previously no such program existed. Our heat mapping software enables the visualization of the retinal thickness across the whole retinal scans facilitating the comparison of specimens and localization of areas of interest. Our novel algorithm provides the first tools to analyze SD-OCT scans of non-mammalian species at a faster rate than manual analysis, increasing the potential of future research output. The adaptability of our programs makes them potentially suitable for the analysis of SD-OCT scans from other non-mammalian species